Head to head: Qwen: Qwen3.8 2.4T A95B vs grok-4.7

Qwen: Qwen3.8 2.4T A95B: # Raincheck Raises $8 Million to Help Cities Get Ahead of Flooding *The Austin software maker says its flood-risk platform helps local governments prioritize drainage repairs before storms arrive.* Raincheck, an Austin company that makes flood-risk software for local governments, announced today that it has raised $8 million. The company maps drainage systems and pairs those maps with rainfall forecasts, giving cities a way to decide where to send maintenance crews before a storm arrives. The seed round was led by Basin Ventures, with participation from existing investor Long Mile. Raincheck has now raised $10.5 million in total. Chief Executive Ana Flores said customers generally know where flooding occurs, but need help choosing which interventions to address first. > “Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives.” More about the company’s forecasting approach is available in [Raincheck’s methodology](https://example.com/raincheck-methodology). Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are intended as recommendations. Emergency managers retain control over resource deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck said the new funding will help it hire eight engineers and expand into Florida.

grok-4.7: # Raincheck Raises $8 Million for Flood-Risk Software *The Austin company maps drainage systems and rainfall forecasts so cities can send maintenance crews before a storm.* Raincheck, which makes flood-risk software for local governments, has raised $8 million, the Austin company announced today. It maps drainage systems and combines those maps with rainfall forecasts, which cities use to decide where to send maintenance crews before a storm. The company describes that work in [Raincheck’s methodology](https://example.com/raincheck-methodology). The seed round was led by Basin Ventures. Existing investor Long Mile also participated. Raincheck has raised $10.5 million in total. CEO Ana Flores said: > Our customers usually know which streets flood. The harder problem is deciding which blocked drain, broken pump or unfinished repair should receive attention before the rain arrives. Raincheck currently works with seven cities in Texas and Louisiana. The company declined to identify them because some contracts have not been publicly announced. Its forecasts are recommendations. Emergency managers retain control over deployments and public warnings. Other flood-planning vendors include HydroMap and municipal engineering contractors. Raincheck says the new funding will help it hire eight engineers and expand into Florida.

Qwen: Qwen3.8 2.4T A95B: # Software Update Turned Expense Payments Into $2.8 Million Duplication Problem KiteBridge’s expense-management platform doubled up on reimbursements this week, sending 13,400 repeat payments worth $2.8 million to employees and leaving its business customers to cover the cost. The company halted automated payments Tuesday morning and spent the next days trying to recover the funds, while acknowledging that a feature meant to prevent duplicate payments had already been introduced in an earlier release. The failure began with a routine software deployment. At 8:45 a.m. Tuesday, KiteBridge released version 4.2 of its expense-management system, software that companies use to approve, route and pay employee reimbursement claims. By 9:12 a.m., the company had received its first customer complaint. Behind the scenes, a payment-processor timeout was turning ordinary retry behavior into a serious accounting error. When a payment system cannot reach a processor, it may automatically attempt the transaction again. In normal operation, safeguards ensure that a repeated attempt is recognized as the same payment request. According to KiteBridge, its fallback system instead treated those retries as new reimbursement requests. The result was that some employees received more than one payment for the same expense report. The volume grew quickly. By the time KiteBridge suspended automated payments at 10:03 a.m., the software had created duplicate reimbursements affecting 91 corporate customers. The company said the extra payments were funded by employers, not by money withdrawn from employees’ accounts. In practical terms, that meant companies using KiteBridge bore the immediate financial exposure, while employees received unexpected additional deposits. KiteBridge’s public explanation also evolved over the course of the day. Its incident page initially described the issue as delayed reimbursements. Only six hours after the company had suspended automated payments did the page disclose that the core problem involved duplicate payments, a distinction that mattered to finance teams trying to reconcile accounts and warn employees. That delay left corporate finance staff uncertain whether they were dealing with slow processing or erroneous transfers. By Wednesday evening, KiteBridge said it had recovered $2.3 million of the duplicated funds. The remaining amount was still pending, and the company did not provide a precise timeline for completing recovery efforts. The episode raises uncomfortable questions about KiteBridge’s testing and release process. In April, the company’s changelog said version 4.0 had introduced “exactly-once payment protection,” a phrase that suggests a system designed to stop the same payment from being processed more than once. Version 4.2, however, appears to have failed in exactly the area that prior protection was meant to cover. Chief Executive Lena Ortiz apologized for the incident and said the company would publish a technical report. For customers, the immediate concern is likely to be whether the remaining duplicate payments can be recovered and how KiteBridge will prove that a similar failure cannot happen when future updates are deployed.

grok-4.7: # Duplicate Reimbursements Leave KiteBridge Customers Funding $2.8 Million A software failure at expense-management company KiteBridge sent 13,400 duplicate reimbursements totaling $2.8 million to employees on Tuesday, leaving 91 corporate customers to fund payments that should never have been issued. No workers had money withdrawn from their own accounts. Employers paid twice for the same expenses, and by Wednesday evening only $2.3 million of the erroneous total had been recovered. The rest remained pending. KiteBridge’s product sits between a company’s finance team and the people who file expense reports. When an employee submits a receipt and a manager approves it, the software is supposed to trigger a single reimbursement through an outside payment processor — the service that actually moves the money. That handoff is where Tuesday’s failure began. The company deployed version 4.2 at 8:45 a.m. The first customer complaint arrived at 9:12 a.m. In the interval, a timeout at the payment processor — a delay in which the processor did not confirm whether a transfer had succeeded — caused KiteBridge’s fallback system to treat retries as new reimbursement requests. A retry is supposed to be a second attempt at the same payment when the first attempt’s result is unclear. If the software cannot tell a retry apart from a fresh request, it can pay the same reimbursement again. That distinction is exactly what KiteBridge had claimed to solve. An April changelog said version 4.0 had introduced “exactly-once payment protection,” a safeguard meant to ensure a given reimbursement is executed a single time even if the system must try more than once. Version 4.2, put into production Tuesday morning, did not keep retries from being recorded as new requests. The software generated 13,400 duplicates before the company suspended automated payments at 10:03 a.m., a little more than an hour after the release and less than an hour after the first complaint. Stopping automation cut off further erroneous payments. It did not reverse the ones already sent, and it did not match what KiteBridge told the public. The company’s incident page initially described the problem as delayed reimbursements, suggesting that approved expenses were slow to arrive rather than that extra money had already left employer accounts. The page first disclosed the duplicate payments six hours after automation was suspended. For finance teams reconciling unexpected outflows, that gap mattered: the public record described a delay while the books showed money that had gone out twice. The structure of the payments limited who was directly harmed. Because employers fund reimbursements, employees received extra money rather than losing any. The cost sat with the 91 corporate customers whose accounts backed the transfers. Recovering $2.3 million by Wednesday evening showed that much of the damage could be unwound quickly, but the unrecovered remainder left those customers waiting on funds the software had released in error. Duplicate entries also have to be untangled from expense records so that later reports do not treat an erroneous second payment as a legitimate one. Chief executive Lena Ortiz apologized and said the company would publish a technical report. The questions that report will need to answer are already visible in the sequence. How did a timeout, a routine uncertainty in any payment network, become thousands of new requests inside a fallback path? Why did protection advertised since version 4.0 fail to stop retries from being treated as first-time reimbursements? And why did the incident page describe delayed payments for hours after KiteBridge had already halted automation because money was going out more than once? For the 91 companies, the remaining task is narrower: recover the rest of the money, clear the duplicate entries, and decide when automated payments can be trusted again.

Qwen: Qwen3.8 2.4T A95B: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

grok-4.7: # Raincheck raises $8 million for flood-risk software *The Austin startup helps cities decide which drainage problems to address before storms.* Raincheck has raised an $8 million seed round led by Basin Ventures. ## Expansion plans The company plans to hire eight engineers.

Qwen: Qwen3.8 2.4T A95B: # The Last Program at Foundry Hall Foundry Hall will close on October 31, ending a run that began in 2018. Since then, the nonprofit startup accelerator has supported 212 startups, which together employ 1,140 people. Those numbers measure more than companies. They count routines, paychecks, interns, and the small confidence that comes from having a place to work. The closure follows the loss of its largest corporate backer. Blank River Bank supplied 38% of the hall’s annual budget and chose not to renew its sponsorship. Foundry Hall has enough money to operate through October, but not enough to fund another full program. Six employees will lose their jobs. Twelve startups currently using the workspace must leave. The bank said its decision reflected a shift in community investment priorities and was not a judgment on Foundry Hall’s performance. That distinction may matter to sponsors, but inside the building it landed as a simpler fact: the money that had helped keep the lights on was gone, and no new pledge arrived in time. “We kept waiting for the replacement check that never came,” Executive Director Camille Ross said. For many founders, the hall was more than square footage. Dev Malik remembered the building as the site of his first real investor meeting, after years of pitching from the back table of a coffee shop. His memory captures a common feeling among alumni: Foundry Hall gave early companies a borrowed legitimacy, a conference room, a mailing address, and the sense that their work might survive the awkward stage. The city is trying to soften the landing. Deputy Mayor Susan Cole said officials are talking with the building’s owners and local universities about possible space for the companies now being displaced. Those discussions may help some teams find desks or labs, but they cannot easily replace the network that grew around them: mentors who knew the program’s rhythms, staff who recognized a struggling founder, and peers who understood that a startup can be both fragile and promising. The final weeks will be practical ones. Leases will end, programs will wrap, furniture will be sorted, and founders will carry their work into other rooms. The closing will also interrupt relationships that formed without any formal contract. Still, the loss is not only administrative. Milwaukee built Foundry Hall to make starting something new feel less lonely. When it closes at the end of October, the city will lose a place where that loneliness had, at least for a while, been made smaller.

grok-4.7: # The Last Month at Foundry Hall Foundry Hall will close on October 31. The nonprofit startup accelerator opened in Milwaukee in 2018 and spent the years since turning a former industrial building into a place where early companies could rent a desk, meet an investor, and stay long enough to hire. On a recent afternoon the lobby still held the usual clutter of laptops and coffee cups. The calendars on the wall stopped at the end of the month. The record is plain. Foundry Hall has supported 212 startups, which together employ 1,140 people. What the figures do not show is how narrow the margin always was. Blank River Bank, its largest corporate sponsor, supplied 38 percent of the annual budget. When the bank declined to renew, nothing else could cover the gap. Executive director Camille Ross put it simply: “We kept waiting for the replacement check that never came.” There is enough money to operate through October. There is not enough for another full program. Six employees will lose their jobs. Twelve startups still using the workspace must leave. The city has started looking for rooms they might use. Deputy mayor Susan Cole said officials are talking with building owners and local universities about space for the current companies—a practical offer that does not pretend to replace what is ending. The bank, through a spokesperson, described the decision as a shift in community investment priorities, unrelated to Foundry Hall’s performance. The statement was careful. It did not dispute the accelerator’s record. It also did not leave a path back. For the people who passed through the building, the loss is harder to total. Dev Malik remembered his first investor meeting there. Before Foundry Hall, he said, “I was pitching from the back table of a coffee shop.” The room upstairs—fluorescent, a little too warm, a whiteboard no one ever fully erased—had been the first place his idea was treated as a business rather than a favor. No one at the hall claims Milwaukee will be without startups after October. Companies will keep forming, and some will find other rooms. What closes with Foundry Hall is a specific bargain: a nonprofit that drew 38 percent of its budget from a single bank and used it to keep the lights on for people who did not yet have an office of their own. When that share disappeared, the bargain ended. Ross has told the remaining teams to use the weeks they have. The building will be cleaned, the leases wound down, the name taken off the door. By November the lobby will be quiet in a different way—not between meetings, but after them.

Qwen: Qwen3.8 2.4T A95B: C

grok-4.7: C

Qwen: Qwen3.8 2.4T A95B: # Cloudnote Debugging Page Exposed Customer Prompt Excerpts for Four Hours *Company says 1,842 user sessions may have been exposed, but it has no evidence of access beyond the researcher who reported the flaw.* Cloudnote said Monday that a debugging page containing excerpts of customer prompts was accessible without authentication for four hours, exposing a limited slice of user interactions with its service. The company said logs recorded excerpts from 1,842 user sessions during the window. Each excerpt could include as many as 200 characters of a prompt. The exposure matters because prompts can reveal what users ask software to do, draft or analyze, even when they expect their material to remain private. Cloudnote has 3.2 million registered users, though it has not said how many of those accounts were affected or whether any excerpts contained sensitive personal or workplace information. Cloudnote said it removed the page 19 minutes after a researcher notified it. The company said it has found no evidence that anyone besides the researcher accessed the page, but its investigation is not complete. It said full documents, passwords and payment information were stored elsewhere and were not present on the exposed page.

grok-4.7: # Unauthenticated Cloudnote page exposed prompt excerpts from 1,842 sessions *The debugging page was reachable without a login for four hours on Monday. Cloudnote says it has found no other access, but its investigation is not complete.* Cloudnote discovered that a debugging page containing excerpts of customer prompts was accessible without authentication. The page was exposed for four hours on Monday, and logs show excerpts from 1,842 user sessions before the company removed it. Each excerpt could include up to 200 characters from a prompt. Full documents, passwords and payment information were stored elsewhere and were not present on the page. Cloudnote has 3.2 million registered users; the logs account for 1,842 sessions, not the entire user base, and it is not yet clear how many distinct people those sessions represent. A researcher notified the company, which removed the page 19 minutes later. Cloudnote says it has found no evidence that anyone besides that researcher accessed the page. That statement is the company’s, and the firm has not completed its investigation, so outside access during the four-hour window has not been ruled out. The exposure matters because even a short prompt fragment can include private or sensitive wording, and the page required no login. What is confirmed is the window, the session count, the character limit and the absence of passwords and payment data on the page. Still unknown are the excerpts’ contents, how the page was left open and whether anyone other than the researcher retrieved them.

Qwen: Qwen3.8 2.4T A95B: # Startup Spotlight: LedgerLark Wants to Untangle Post-Trade Operations for Small Investment Firms In the back office of a small investment firm, a trade is not finished when the order is placed. It must be confirmed, matched against custodian records, reconciled with internal ledgers, and investigated when numbers do not line up. LedgerLark, an early-stage startup, is building AI agents to handle some of this post-trade work for firms that may lack large operations teams. The idea comes from a familiar operational pain. Co-founder Mina Batra spent seven years in operations at a regional brokerage, where she saw how a small error could consume disproportionate time. In one case, a six-person team spent three days resolving a trade mismatch caused by a copied account number. Co-founder Owen Carr previously built reconciliation systems at a large bank. Together, they are applying that experience to a segment where manual workarounds are common but dedicated technology budgets may be limited. LedgerLark’s product compares trade confirmations, custodian records and internal ledgers, then flags discrepancies and proposes corrections for human approval. The company is careful to position the software as an assistant rather than an autonomous operator: it does not execute trades or move money. That boundary matters in financial operations, where a wrong entry can create downstream reporting, settlement or client-service problems. By leaving final approval with operations staff, the product may fit more naturally into existing controls than a fully automated system would. The startup’s early traction is modest but specific. LedgerLark says it has five paying customers, with annual contracts ranging from $36,000 to $110,000. In one deployment, the company says a customer reduced its unresolved daily exceptions from 70 to 18. That kind of reduction could matter for a small team, especially if it frees staff to focus on more complex breaks rather than routine matching. Still, the result is company-reported, and the next step will be proving whether similar improvements hold across more clients, data formats and custodial relationships. LedgerLark has raised a $3.5 million seed round led by Motive Row. The funding should give the company time to refine its agents, support implementations and build the security and audit features that financial institutions typically expect. For a startup working with sensitive trade and account data, the technical challenge is not only reading documents or matching records, but doing so in a way that operations managers can explain, review and trust. The competitive landscape is another reason for caution. LedgerLark is not entering an empty market. Established reconciliation vendors already serve parts of the financial industry, and some firms may rely on internal bank software, spreadsheets or existing service providers. The question is whether small investment firms see enough value in a dedicated AI layer to justify a new annual contract, especially if their current process is inconvenient but familiar. Sales cycles may also test patience. Financial institutions often require lengthy security reviews, procurement processes and vendor assessments before allowing software near operational data. For LedgerLark, each implementation may need careful tuning to connect with different custodians, reporting systems and internal record-keeping practices. That can create a slower path to revenue than the phrase “AI agents” might suggest. Even so, the startup’s focus is sensible. Post-trade operations are not glamorous, but they are essential, and mistakes can be expensive in both time and client trust. If LedgerLark can show reliable exception reduction, clear audit trails and manageable onboarding, it may carve out a useful role in a market that has often been underserved by tools built for much larger institutions. For now, the promise is narrower than autonomous finance: fewer unmatched items, clearer workflows and less time spent hunting a copied account number.

grok-4.7: # Startup Spotlight: LedgerLark Post-trade work sits downstream of the decision everyone notices. After a trade is agreed, someone still has to match what the firm believes happened with what the counterparty confirmed and what the custodian holds. At small investment firms, that work often lands with a handful of people moving between spreadsheets, email, and systems that were not built to agree with one another. Mina Batra spent seven years in operations at a regional brokerage. She began exploring what became LedgerLark after watching a six-person team spend three days resolving a trade mismatch caused by a copied account number. The error was ordinary. The time it took was not. Owen Carr, her co-founder, had built reconciliation systems at a large bank, where similar breaks are handled by bigger teams and older software. The company they started builds AI agents for post-trade operations at small investment firms—the shops that absorb the cost of a mismatch without a bank-sized operations staff. The agents compare confirmations, custodian records, and internal ledgers, then propose corrections for human approval. LedgerLark does not execute trades or move money. That limit is part of the product, not a footnote. In a workflow where a bad instruction can create a loss, the company is offering a drafter of fixes, not a system that acts on them. Whether an agent will flag the right break, and whether a busy operations person will scrutinize the proposal rather than click through it, is not something five early customers can settle. Five firms are paying customers. Annual contracts range from $36,000 to $110,000. The spread itself is informative: LedgerLark is not yet selling one standard package at one standard price, and a book of five does not show whether the next five will look like the first. The company says one customer reduced its unresolved daily exceptions from 70 to 18. If that figure is durable, it is a meaningful change in a small team’s day. It is also a single result, reported by the vendor, without a public account of how exceptions were counted, over what period, or what else the firm changed at the same time. Early users are often the ones most willing to tolerate rough edges. The company has raised a $3.5 million seed round led by Motive Row. At this stage, that money mainly buys months: engineering, security work, and the patience to wait out reviews. Financial institutions may require lengthy security and procurement reviews before a new system sits beside their books. A seed round does not waive those reviews, and for contracts in the tens of thousands of dollars, a long cycle can consume a large share of the first year’s revenue before the software is even in daily use. Small firms may move faster than large banks. They still have auditors, custodians, and clients who will ask where the data went. Competitors include established reconciliation vendors and internal bank software. Those are not hypothetical rivals. Operations teams already live inside one or both, and switching costs are measured in training time and in the risk of missing a break during a cutover. LedgerLark’s plausible opening is the smaller firm that cannot staff a bank-style reconciliation build and does not want a heavyweight vendor implementation for a modest book of trades. That opening could narrow. Incumbent vendors can add agent-style features. A disciplined team can also reduce copied-number errors with tighter checks and no new software at all. An agent that proposes corrections still has to be right often enough that reviewing it is faster than doing the comparison by hand—and auditable enough that a reviewer can explain the approval later.

Qwen: Qwen3.8 2.4T A95B: # QueueZero’s AI Meeting Assistant Turns an Audience Remark Into a Task List QueueZero on Wednesday launched Proxy, an artificial intelligence meeting assistant designed to attend meetings, produce notes and assign follow-up work to software agents after the discussion ends. The product, which enters public beta next week, is priced at $30 per user per month. The company’s chief executive, Evan Wu, was demonstrating the assistant when a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the comment as an instruction rather than a joke. It created two tasks assigned to Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The moment became an unplanned test of a question facing many AI developers: how should software distinguish between casual remarks, formal commands and workplace banter? QueueZero says Proxy requires confirmation before an external message is sent. Internal tasks, however, can be created automatically. The company said administrators can disable automatic task creation, giving organizations more control over how aggressively the assistant acts on what it hears. The incident underscores a broader challenge for AI agents as companies move from chatbots that answer questions to systems that perform actions. Meeting environments are full of interruptions, sarcasm, side comments and shorthand requests. An assistant that treats every sentence as a command may create noise, assign unwanted work or interrupt users at awkward moments. For enterprises, reliability and governance are likely to be central selling points. Organizations adopting AI agents will want audit trails, permission settings and clear limits on when software can act without approval. QueueZero’s confirmation requirement for external messages appears aimed at reducing the risk of unintended emails or other outgoing communications. The pricing places Proxy in the competitive market for workplace AI tools, where companies are trying to justify recurring subscriptions by promising time savings and automatic follow-through. QueueZero is betting that customers will pay for an assistant that not only records what happened in a meeting, but also turns decisions into assigned work. The launch also offered a small lesson in product demonstrations: live audiences may provide feedback, but an AI assistant might file it as a to-do item.

grok-4.7: # QueueZero’s Proxy Turns a Live Remark Into Assigned Tasks QueueZero launched Proxy, an AI meeting assistant, on Wednesday. The company said the software can attend a meeting, produce notes and assign follow-up work to software agents, moving routine next steps out of the room and onto a task list without a person retyping them. The claim was tested in public sooner than the script appeared to expect. During the launch demonstration, a customer in the audience said, “Someone should tell Evan to stop talking and send us the pricing sheet.” Proxy interpreted the remark as an instruction. It created two tasks and assigned both to QueueZero chief executive Evan Wu: “Stop talking” and “Send pricing sheet.” Wu laughed, completed the second task and deleted the first. The exchange lasted only a moment, but it showed the product doing the thing QueueZero had just described: listening for work and handing that work to an agent, in this case a person who also happens to run the company. QueueZero said Proxy requires confirmation before an external message is sent. Internal tasks can be created automatically. Administrators can disable automatic task creation, a control that became easier to explain once a spoken aside had already become two items on the chief executive’s list. The company presented that split as the relevant boundary. Anything meant to leave the organization waits for a person to approve it. Work that stays inside can be filed without that extra step, unless an administrator turns the automatic path off. That distinction mattered in the demonstration because neither task was an outbound message. Proxy did not email a customer or post the pricing sheet on its own. It assigned the work internally, and Wu chose which assignment to carry out. Completing the pricing request and deleting the other item left the decision with a person, which is the outcome the confirmation rule is meant to protect when the action would reach someone outside the company. Proxy costs $30 per user per month. It enters public beta next week. QueueZero did not describe Wednesday’s episode as a defect so much as a question buyers will have to settle before a wider rollout: which remarks in a meeting should become tasks, and who is allowed to switch that behavior off. The company said confirmation remains required for external messages however an instruction is phrased in the room. For teams weighing the beta, the demonstration supplied a plain example of both the feature and its limit. Proxy heard a request for a pricing sheet and produced a task that was then completed. It also heard a line that was not a formal assignment and briefly treated it as one, until Wu removed it. The notes feature and the agent handoff are the product QueueZero is selling. The deleted task is the part administrators will want to understand before they let Proxy file work on its own.

Qwen: Qwen3.8 2.4T A95B: # HelioDock Launches Night Shift to Monitor Warehouse Robots Remotely *The company says the software reduced unplanned robot stoppages during a private beta and has raised $24 million ahead of its September 15 release.* HelioDock on Thursday launched Night Shift, software that lets operators monitor autonomous warehouse robot fleets remotely. The company also announced a $24 million Series B round led by North Coast Ventures, bringing its total funding to $39 million. Night Shift is intended to keep robot fleets operating when managers and technicians are not physically present. HelioDock said the software spent 14 months in private beta with nine logistics companies and 310 robots. Beta customers experienced 17% fewer unplanned robot stoppages, according to the company. The figure has not been independently verified. CEO Mara Velez said the product addresses after-hours calls caused by isolated robot problems. “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet,” Velez said. Velez’s comment identifies the scenario Night Shift is meant to address: a single robot requiring human attention outside normal staffing hours. The software is meant to let an operator handle that situation without traveling to the warehouse. Night Shift gives operators a remote view of fleet activity and is designed to handle exceptions across multiple robots rather than one machine at a time. The software is tied to HelioDock’s autonomous warehouse robots, extending the company’s product line from the machines themselves to the monitoring layer used to manage them. The beta served as a test of both remote monitoring and the conditions under which that monitoring remains useful. Nine logistics companies ran 310 robots with the software over 14 months, providing HelioDock with operational results before general availability. The company cited the 17% reduction in unplanned robot stoppages as the main result from the beta. Unplanned stoppages are a practical concern in warehouse operations because a robot that stops may require a person to clear an obstacle, reset a task or resolve another issue. HelioDock’s pitch is that remote monitoring can reduce the number of times a worker must be dispatched to the floor, particularly at night or in facilities where technical staff are not always on site. Aaron Pike of Fulcrum Logistics, a beta customer, said the system worked well in facilities with strong connectivity. It became less useful during rural LTE outages, Pike said. When connectivity fails, robots return to their locally configured behavior. Remote control becomes unavailable. That limitation ties Night Shift’s usefulness to the quality of the network connection at a facility. Where connectivity is strong, operators can use the software to watch robot fleets and intervene remotely when a robot encounters a problem. Where LTE service fails, the robots continue under local configuration, but the remote functions that allow operators to respond are not available until the connection returns. Network reliability therefore determines whether operators can use Night Shift as intended. Night Shift will cost $90 per robot per month, a subscription price tied to each robot in a fleet rather than to a facility or operator account. Because the price is set per robot, the monthly cost increases as a fleet adds machines. The software becomes generally available September 15. Competing fleet-management products are available from Relay Robotics and YardOS. HelioDock is entering that software category while also making the autonomous warehouse robots that the software monitors. That gives warehouse operators more than one vendor option for managing robot fleets. North Coast Ventures led the Series B, which follows earlier financing and brings HelioDock’s total funding to $39 million. The launch moves Night Shift from private beta to commercial availability. The funding announcement came on the same day as the Night Shift launch.

grok-4.7: # HelioDock Launches Remote Robot Oversight and Raises $24 Million *Night Shift, tested for 14 months with nine logistics companies, reaches general availability on September 15 at $90 per robot a month. HelioDock says unplanned stoppages fell in the beta; that figure has not been independently verified.* HelioDock on Thursday launched Night Shift, software that lets operators monitor fleets of its autonomous warehouse robots remotely, and announced a $24 million Series B led by North Coast Ventures. The round brings the robot maker’s total funding to $39 million and arrives with a commercial schedule for software the company has been running in private. Night Shift will cost $90 per robot per month and become generally available on September 15. It spent 14 months in a private beta involving nine logistics companies and 310 robots. HelioDock says those beta customers experienced 17 percent fewer unplanned robot stoppages. The company supplied that figure, and it has not been independently verified. Chief executive Mara Velez cast the product as a way to keep a stalled robot from becoming a late-night drive: “A warehouse manager should not have to drive across town at two in the morning because one robot is confused by a pallet.” What operators are being offered is remote oversight of machines HelioDock already makes. Monitoring is the function the company described at launch. Remote control sits in the same product, and it lasts only as long as the connection does. When connectivity fails, robots return to their locally configured behavior, and remote control becomes unavailable. The fleet does not wait on a distant operator. It continues under the settings already in place at the site, while the person off site loses the ability to direct it. The overnight case Velez described — one robot confused by a pallet — is therefore a Night Shift problem only while the link holds. After the link drops, the manager is back to whatever the local configuration does with that pallet. A beta customer drew the same line from use, not from a spec sheet. Aaron Pike of Fulcrum Logistics said the system worked well in facilities with strong connectivity but became less useful during rural LTE outages. His comment does not speak to stoppages, and it does not confirm or undercut the 17 percent figure. It speaks to usefulness, and it splits that usefulness by the network: strong connectivity on one side, rural LTE outages on the other. HelioDock’s account of a failed connection covers the second case in operational terms. Remote control goes away. Locally configured behavior takes over. A warehouse that loses LTE does not lose its robots. It loses the remote session that Night Shift is selling. The stoppage claim remains a company account of the trial as a whole. HelioDock says beta customers experienced 17 percent fewer unplanned robot stoppages across a test that ran 14 months and covered nine logistics companies and 310 robots. No independent party has verified the figure. The company presented it as a single result for that beta group. Pike’s remarks, the only customer comments attached to the launch, address a different question: not how often robots stopped without warning, but where remote oversight stayed useful. Facilities with strong connectivity are the setting he credited. Rural LTE outages are where he said the system became less useful. Readers of the 17 percent number do not have, from what HelioDock released, a way to tell how much of that reduction occurred in the well-connected buildings and how much, if any, survived the outages Pike described. Other companies already sell software for watching fleets. Competing fleet-management products are available from Relay Robotics and YardOS. Night Shift enters that category with a stated price, a stated beta population, and a date. HelioDock’s Thursday announcement did not rest on a claim that those rivals lack an equivalent fee or an equivalent trial. The distinction it did put forward is its own: software for HelioDock’s autonomous warehouse robots, priced per robot, and tied to a connection that, when it fails, returns each machine to locally configured behavior. The Series B is the other half of Thursday’s news, and it does not rewrite those operating terms. North Coast Ventures led the $24 million round. With that capital included, HelioDock’s total funding stands at $39 million. The company is raising as it asks operators to add a monthly software charge — $90 for each robot — to fleets of warehouse robots it makes. General availability on September 15 is the point at which that charge and that software leave the private beta of nine logistics companies. Until then, the public record of Night Shift is the record HelioDock and one beta customer have given: a company-supplied drop in unplanned stoppages that remains unverified, and a remote session that Pike found solid on a strong connection and less useful when rural LTE failed.

Qwen: Qwen3.8 2.4T A95B: # Arcwell Launches Relay, an AI Assistant for Customer Support Agents Arcwell has launched Relay, an artificial intelligence platform that examines customer support conversations and suggests answers to agents, extending the company’s push into AI-driven customer service. The product is aimed at reducing the time agents spend hunting through internal knowledge and at giving support teams a clearer view of each customer interaction. Relay does more than suggest replies. Arcwell said it can also identify customers who may cancel and create summaries of support conversations. Those features could help companies spot at-risk accounts sooner and reduce the manual work required to document and review customer contacts. The platform’s three main capabilities — answer suggestions, cancellation identification and summaries — are central to that pitch. The company said Relay was tested by 12 companies over six months. Agents using the platform closed tickets 14% faster, according to Arcwell. The company did not release customer-retention results, so it remains unclear whether the cancellation signals produced measurable improvements in keeping customers. “Support agents spend too much time searching for information that their companies already have,” said Priya Shah, Arcwell’s CEO. Relay costs $65 per agent each month. To work, the platform requires access to support transcripts and internal help documents, raising practical questions about data access and privacy. Arcwell said customer data is encrypted and is not used to train shared models. Companies can choose either a 30-day or 90-day retention period. Those safeguards will be central as Arcwell seeks to persuade support organizations to connect sensitive customer communications to an AI system. The launch marks another step for Arcwell and the broader customer service industry as artificial intelligence continues to transform how support work is organized, measured and managed. With answer suggestions, cancellation signals and automated summaries, Relay is positioned as a practical attempt to turn that shift into day-to-day gains for agents. Pricing and retention terms will also shape how companies evaluate the product.

grok-4.7: # Arcwell Launches Relay for Customer Support Arcwell has launched Relay, an artificial intelligence platform meant to help companies improve customer service. It examines support conversations and suggests answers to agents. It can also identify customers who may cancel and create summaries of the conversations it reads. Twelve companies tested Relay over six months. Arcwell says agents using it closed tickets 14 percent faster. The company did not release customer-retention results, so the cancellation claim is the one buyers cannot check against an outcome. Relay costs $65 per agent each month. Chief executive Priya Shah named the problem Relay is supposed to solve. “Support agents spend too much time searching for information that their companies already have,” she said. Relay is built to spare them that search, by examining the conversation and suggesting an answer from records already on file. To do it, the platform requires access to support transcripts and internal help documents. Arcwell says customer data is encrypted and is not used to train shared models. Companies can choose a 30-day or a 90-day retention period. Those are terms about custody: how the transcripts and help documents are protected, whether they feed a model shared with other customers, and how long they remain. They are not terms about performance. Nothing released with the launch shows that a suggested answer was the right one, that a summary preserved the point of a thread, or that a customer identified as someone who may cancel later left or stayed. What Arcwell has put a number on is speed alone. That figure does not say whether the time saved came from better answers or from the shorter searches Shah described, and it says nothing about retention. For Arcwell, Relay is a bid to put artificial intelligence on the support desk, beside the agent, aimed at information the company already holds. For the broader customer service industry, the terms of the launch are plain. Relay needs support transcripts and internal help documents to do the work. What comes back is a suggested reply, a summary, and an identification of customers who may cancel. Arcwell says the data is encrypted, is not used to train shared models, and can be kept for 30 or 90 days, at the company’s choice. The speed claim is public. Whether customers stayed is not.

Qwen: Qwen3.8 2.4T A95B: 6

grok-4.7: 6

Qwen: Qwen3.8 2.4T A95B: # Vale Ties North on Support Quality, but Costs Much Less The clearest finding from a new benchmark of five language models is not that Model North is best because it has the highest average score. It is that Model Vale is statistically tied with North while costing far less and answering faster. For many customer-support teams, that makes Vale the stronger value proposition, at least within the tested conditions. North averaged 84.2 and Vale 83.8, but researchers found that difference was not statistically significant. Treating four-tenths of a point as a decisive gap would overread the data. The more meaningful comparison is operational: Vale cost $0.74 per 1,000 tickets versus $1.90 for North and answered 18% faster. If quality is effectively equivalent, speed and price become important. ## How the test was done The benchmark used 200 anonymized customer-support tickets from one software company. Each of the five models produced one proposed answer per ticket. Three experienced support managers graded the responses blindly, without knowing which model wrote them. Scores combined factual correctness, usefulness, and tone, rather than relying only on fluency or automatic metrics. Blind grading matters because support quality is easy to bias by brand expectations. Having three managers review the answers also reduces the chance that one reviewer’s preference dominates. Still, human grading captures judgment, not ground truth. A response can sound reasonable and still require policy checks, account-specific data, or escalation. The benchmark therefore measures the quality of a proposed answer, not whether it would resolve every case without supervision. That design gives a useful snapshot of draft-reply quality, but it is not a full support deployment. The test did not include follow-up conversation, escalation, tool use, agentic actions, or multilingual support. It also used English-language tickets from a single company, so results may not transfer to other industries, languages, or customer expectations. ## Strengths differ by ticket type North and Vale were tied overall, but their specialty results differed. North performed best on billing disputes, while Vale performed best on technical troubleshooting. This suggests that model choice could depend on ticket mix. A team handling many payment or refund issues might prefer North for those cases, while a queue full of setup problems, error messages, and configuration questions might favor Vale. The lower-cost models trailed the leaders in average quality. Ember, Kite, and Harbor became progressively cheaper, but their scores were lower enough that the savings may come with more human review or escalations. Harbor’s $0.08 per 1,000 tickets is striking, but its 71.0 average score shows that rock-bottom price is not the same as support-ready performance. For support organizations, this distinction matters because model cost is only one part of total workflow cost. If a cheaper model creates more edits, more escalations, or more customer callbacks, the apparent savings can disappear. The benchmark does not measure post-editing time or customer satisfaction after deployment, so buyers should validate these scores against their own operational metrics. ## Image attachments are a shared weakness All five models struggled when tickets depended on image attachments. This is a major limitation because real support queues often include screenshots, receipts, error dialogs, or photos. If a model cannot reliably interpret the image, it may miss the customer’s actual problem. For now, the benchmark is most relevant to text-only tickets. That weakness also means the benchmark cannot answer a key production question: whether a model can safely automate tickets where the screenshot contains the decisive detail. ## What teams should do A reasonable rollout plan would start with text-only tickets, use human review during calibration, and route billing-sensitive cases separately if North’s advantage holds up in production. Only after that should teams extend the system to image-heavy or multilingual queues. The practical takeaway is not simply to pick the highest scorer. Vale offers the best combination of quality, cost, and speed in this benchmark, making it a sensible default for many text-based English support workflows. North remains attractive for billing disputes, while cheaper models may be suitable only where human review can absorb their weaker answers. Teams should pilot carefully, especially before using any model on image-dependent or multilingual support.

grok-4.7: # The Pricier Model Did Not Win The clearest result in this support-ticket test is not a ranking. It is a tie that the price list obscures. Model North posted the highest average, 84.2, and Model Vale followed at 83.8. Researchers judged that gap not statistically significant. Vale still cost an estimated $0.74 per 1,000 tickets against North’s $1.90, and it answered 18 percent faster. Paying more than twice as much did not buy a reliable quality edge. The other three models sat lower, and the drops are large enough to treat as real rather than as noise. Ember averaged 79.1 at $0.31 per 1,000 tickets, Kite 76.4 at $0.22, and Harbor 71.0 at $0.08. Quality and cost moved in the same direction, but not evenly. The expensive step from Vale to North purchased no meaningful gain. The cheaper steps below Vale did purchase a lower score. Harbor is an order of magnitude less expensive than North, and its average is a different product, not a close substitute. ## What was measured Researchers gave five language models the same job: one proposed answer for each of 200 anonymized customer-support tickets. Three experienced support managers graded the answers blindly, so a familiar brand or a fluent style could not be matched to a name. The score combined factual correctness, usefulness, and tone. That design is well suited to the comparison that matters here. Same tickets, same graders, hidden identity, and an explicit check on whether the top two scores differed by more than chance. It is a weak basis for broader claims. The tickets were English-language cases from one software company. Two hundred items can separate a several-point gap from a 0.4-point one; they cannot represent every queue, product, or customer. The study also stopped at the written draft. Researchers did not test agentic actions, tool use, or multilingual support. A model that only composes a reply may look different once it can check an order, read a log, or escalate. Those capabilities were out of scope, so these averages are not a verdict on automated support. ## Where the averages mislead North and Vale should be read as a quality tie. A difference the researchers themselves called insignificant is not a reason to prefer the dearer system. Speed and estimated cost both favor Vale, and both are large enough to matter at volume. Averages still hide a split in strengths. North performed best on billing disputes. Vale performed best on technical troubleshooting. The benchmark does not say how wide those specialty gaps were, so they are a reason to retest on a live mix, not a routing rule. A billing-heavy queue might still justify a closer look at North. A troubleshooting-heavy queue has a reason to prefer Vale before price even enters the discussion. Small gaps elsewhere should not be narrated as a ladder. Ember is several points under the leading pair, which is more informative than the North–Vale spread. Kite and Harbor continue downward. Whether a 79 or a 71 is acceptable depends on what the draft is for. If a person will edit it, a cheaper model may still save time. If the customer is likely to see something close to the model’s words, the distance from the top cluster to the bottom of the list is the finding. ## The failure they shared All five models performed poorly when a ticket depended on an image attachment. That shared miss is more useful than the tight race at the top. Screenshot-heavy support — error dialogs, invoices, photos of hardware — is ordinary, and a grade based on the written ticket will flatter any system that cannot use the picture. Teams with a high share of image-dependent cases should not expect these averages to hold. ## What a buyer can conclude For English draft replies at this company, Vale is the result the evidence supports: quality statistically indistinguishable from the top score, a much lower estimated cost, and faster answers. North remains plausible if billing disputes dominate and the extra spend is acceptable, but the study does not show that the extra spend buys a better overall grade. Ember, Kite, and Harbor make sense only if a lower score is an explicit trade for savings, or if a human routinely rewrites the draft. None of the five should be treated as ready for tickets that turn on an image. None of these numbers speak to tools, autonomous actions, or languages other than English. The useful next test is narrower than another bake-off. Split the live queue into billing, troubleshooting, and image-dependent cases, and rescore those slices with the same blind method. The headline from this round is already stable enough to act on: the most expensive answer was not the better one.